AI coding agents need checks, not just memory
can repeat the same kind of mistake even after new instructions are added. The weak point is that the rule stays in instead of becoming something that blocks risky action. A more reliable pattern is to record the failed action, turn it into a specific pre-action check, and inspect the next before it runs.
If the next action matches a risky pattern, the agent should provide evidence or ask for approval. For example, a protected dashboard page should not show a 404 error when someone is signed out. The better test is to request the protected page with no session, check that it redirects to sign-in, and confirm the return URL is kept.
ThumbGate uses this idea by turning feedback into a local, visible gate before the next , file edit, browser action, deploy, or API call runs.
Key points
- Adding more instructions may not stop an from repeating a mistake.
- A failed action should become a specific pre-action check.
- Risky should be inspected before they run.
- Some actions should require evidence or human approval before execution.
- Repeated failures are better handled as visible gates than as prompt rules.